CN112905876B - Information pushing method and device based on deep learning and computer equipment - Google Patents

Information pushing method and device based on deep learning and computer equipment Download PDF

Info

Publication number
CN112905876B
CN112905876B CN202010181488.8A CN202010181488A CN112905876B CN 112905876 B CN112905876 B CN 112905876B CN 202010181488 A CN202010181488 A CN 202010181488A CN 112905876 B CN112905876 B CN 112905876B
Authority
CN
China
Prior art keywords
information
feature
sequence
user behavior
user
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Active
Application number
CN202010181488.8A
Other languages
Chinese (zh)
Other versions
CN112905876A (en
Inventor
肖小粤
曹星忠
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Tencent Technology Shenzhen Co Ltd
Original Assignee
Tencent Technology Shenzhen Co Ltd
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Tencent Technology Shenzhen Co Ltd filed Critical Tencent Technology Shenzhen Co Ltd
Priority to CN202010181488.8A priority Critical patent/CN112905876B/en
Publication of CN112905876A publication Critical patent/CN112905876A/en
Application granted granted Critical
Publication of CN112905876B publication Critical patent/CN112905876B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/90—Details of database functions independent of the retrieved data types
    • G06F16/95—Retrieval from the web
    • G06F16/953—Querying, e.g. by the use of web search engines
    • G06F16/9535—Search customisation based on user profiles and personalisation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
    • Y02D10/00—Energy efficient computing, e.g. low power processors, power management or thermal management

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Physics & Mathematics (AREA)
  • Databases & Information Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Evolutionary Computation (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Computational Linguistics (AREA)
  • Mathematical Physics (AREA)
  • Biomedical Technology (AREA)
  • Health & Medical Sciences (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Biophysics (AREA)
  • Software Systems (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Evolutionary Biology (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

The application relates to an information pushing method, device and computer equipment based on deep learning, wherein the method comprises the following steps: candidate item information is obtained, and historical behavior information corresponding to a user identifier is obtained; extracting a user behavior sequence corresponding to each behavior type in the historical behavior information; inputting the candidate item information and the user behavior sequence into a prediction model, extracting user behavior sequence features corresponding to the user behavior sequence and item feature vectors of the candidate item information, and determining predicted values of the candidate item information according to the user behavior sequence features and the item feature vectors; screening target pushing information according to the predicted value of the candidate item information; and pushing the target pushing information to the user terminal corresponding to the user identifier. The scheme that this application provided can effectively improve the accuracy of information push.

Description

Information pushing method and device based on deep learning and computer equipment
Technical Field
The present disclosure relates to the field of computer technologies, and in particular, to an information pushing method and apparatus based on deep learning, and a computer device.
Background
With the rapid development of internet technology, more and more users like to read internet books, and users usually select interested information to read aiming at massive various books. With the development of artificial intelligence technology, in order to improve the information pushing efficiency, some information pushing modes aiming at the interests of users are presented.
Conventional information recommendation methods generally include a collaborative filtering-based recommendation method and a content-based recommendation method, wherein the collaborative filtering-based recommendation method is used for recommending by analyzing similar users, and the content-based recommendation method is used for identifying reading interests of the users by analyzing historical access records of the users so as to push related information to the users. However, for newly released information, it is difficult to accurately identify the interests of the user, and there is a problem that the accuracy of pushing the target information is not high.
Disclosure of Invention
Based on the foregoing, it is necessary to provide an information pushing method, an apparatus, a computer readable storage medium and a computer device based on deep learning, aiming at the technical problem that the accuracy of target information pushing is not high.
An information pushing method based on deep learning comprises the following steps:
candidate item information is obtained, and historical behavior information corresponding to a user identifier is obtained;
Extracting a user behavior sequence corresponding to each behavior type in the historical behavior information;
inputting the candidate item information and the user behavior sequence into a prediction model, extracting user behavior sequence features corresponding to the user behavior sequence and item feature vectors of the candidate item information, and determining predicted values of the candidate item information according to the user behavior sequence features and the item feature vectors;
screening target pushing information according to the predicted value of the candidate item information;
and pushing the target pushing information to the user terminal corresponding to the user identifier.
An information push device based on deep learning, the device comprising:
the information acquisition module is used for acquiring candidate item information, acquiring historical behavior information corresponding to a user identifier, and extracting user behavior sequences corresponding to various behavior types in the historical behavior information;
the information prediction module is used for inputting the candidate item information and the user behavior sequence into a prediction model, extracting user behavior sequence characteristics corresponding to the user behavior sequence and item characteristic vectors of the candidate item information, and determining predicted values of the candidate item information according to the user behavior sequence characteristics and the item characteristic vectors;
The information extraction module is used for screening target pushing information according to the predicted value of the candidate item information;
and the information pushing module is used for pushing the target pushing information to the user terminal corresponding to the user identifier.
A computer-readable storage medium storing a computer program which, when executed by a processor, causes the processor to perform the steps of the deep learning-based information push method described above.
A computer device comprising a memory and a processor, the memory storing a computer program which, when executed by the processor, causes the processor to perform the steps of the deep learning based information push method described above.
According to the deep learning-based information pushing method, device and computer equipment, after candidate item information and the historical behavior information corresponding to the user identification are obtained, the user behavior sequences corresponding to the behavior types in the historical behavior information are extracted, so that heterogeneous behavior sequences of the user in various behavior types can be effectively obtained. The candidate item information and the user behavior sequence are input into the prediction model, the item feature vector of the candidate item information and the user behavior sequence feature corresponding to the user behavior sequence are extracted, and then the prediction value of each candidate item information is determined according to the user behavior sequence feature and the item feature vector by using the prediction model based on deep learning, so that the prediction value of each candidate item information can be accurately and effectively output. Target pushing information is screened according to the predicted value of each candidate item information, the target pushing information is pushed to a user terminal corresponding to the user identifier, and the heterogeneous behavior sequence of the user and the candidate item information are learned, so that the multi-mode user interests can be accurately identified, and the accuracy of target information pushing is effectively improved.
Drawings
FIG. 1 is an application environment diagram of an information push method based on deep learning in one embodiment;
FIG. 2 is a flow chart of an information push method based on deep learning in one embodiment;
FIG. 3 is a flowchart illustrating a step of calculating predicted values of a plurality of candidate item information according to one embodiment;
FIG. 4 is a flow chart illustrating a feature combining process for multiple user behavior sequences in one embodiment;
FIG. 5 is a flowchart illustrating a step of extracting target push information in one embodiment;
FIG. 6 is a flow diagram of the steps for training a predictive model in one embodiment;
FIG. 7 is a block diagram of a predictive model in one embodiment;
FIG. 8 is a block diagram of an information pushing device based on deep learning in one embodiment;
FIG. 9 is a block diagram of an information pushing device based on deep learning in another embodiment;
FIG. 10 is a block diagram of a computer device in one embodiment.
Detailed Description
In order to make the objects, technical solutions and advantages of the present application more apparent, the present application will be further described in detail with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are for purposes of illustration only and are not intended to limit the present application.
Fig. 1 is an application environment diagram of an information push method based on deep learning in one embodiment. For example, referring to fig. 1, the deep learning-based information push method is applied to an information push system. The information push system comprises a user terminal 102 and a server 104. The user terminal 102 and the server 104 communicate via a network connection. After obtaining the candidate item information and the historical behavior information corresponding to the user identifier, the server 104 extracts the user behavior sequences corresponding to the behavior types in the historical behavior information. And inputting the candidate item information and the user behavior sequence into a prediction model, extracting item feature vectors of the candidate item information and user behavior sequence features corresponding to the user behavior sequence, and determining the predicted value of each candidate item information according to the user behavior sequence features and the item feature vectors. The server 104 screens the target pushing information according to the predicted value of the candidate item information, and then pushes the target pushing information to the user terminal 102 corresponding to the user identifier. The user terminal 102 may be a desktop terminal or a mobile terminal, and the mobile terminal may be at least one of a mobile phone, a tablet computer, a notebook computer, and the like. The server 104 may be implemented as a stand-alone server or as a server cluster of multiple servers.
As shown in fig. 2, in one embodiment, a deep learning-based information push method is provided. The present embodiment is mainly exemplified by the application of the method to the server 104 in fig. 1. Referring to fig. 2, the information push method based on deep learning specifically includes the following steps:
step 202, obtaining candidate item information and obtaining historical behavior information corresponding to a user identifier.
The candidate item information may be a plurality of pieces of information to be pushed currently recalled by the server. For example, the candidate item information may include article information of a plurality of categories, news information of a plurality of categories, video information of a plurality of categories, and the like. The candidate item information comprises information to be pushed, which is interested by the user.
The historical behavior information of the user refers to behavior information such as requirement expression, information acquisition, information utilization and the like which are expressed when the user acquires required information, and can comprise various behavior types such as clicking behavior information, praying behavior information, comment behavior information, sharing behavior information and the like of push information.
Specifically, the server may acquire a plurality of candidate item information from the information push platform, and the candidate item information may be item information generated within a preset period of time. The server obtains the historical behavior information of the user according to the user identification, wherein the server can obtain the historical behavior information of the user from the local platform and can also obtain the historical behavior information associated with the user from the third-party platform. The server may also obtain historical behavior information of the user from the log information.
Step 204, extracting a user behavior sequence corresponding to each behavior type in the historical behavior information.
The user behavior sequence may be a user behavior based on a time sequence, which represents each step of behavior of the user to engage in a certain activity recorded according to a time sequence in a certain time period. For example, a record of each step of the user's behavior from accessing the website to leaving the website is recorded as a sequence of user behaviors. A user behavior sequence may include information such as a history item sequence, a user identifier, an action behavior identifier, and an operation time.
The historical behavior information of the user comprises user behavior information of a plurality of behavior types. After the server acquires the historical behavior information of the user, the user behavior sequences corresponding to the behavior types in the historical behavior information are extracted. Specifically, the server may identify a behavior type corresponding to each piece of historical behavior information, and extract a user behavior sequence corresponding to a plurality of behavior types from the historical behavior information of the user according to the behavior type. The extracted user behavior sequence is a heterogeneous behavior sequence of the user comprising a plurality of behavior types.
Step 206, inputting the candidate item information and the user behavior sequence into a prediction model, extracting the user behavior sequence characteristics corresponding to the user behavior sequence and the item characteristic vectors of the candidate item information, and determining the predicted value of each candidate item information according to the user behavior sequence characteristics and the item characteristic vectors.
Machine Learning (ML) is a multi-domain interdisciplinary, involving multiple disciplines such as probability theory, statistics, approximation theory, convex analysis, algorithm complexity theory, etc. It is specially studied how a computer simulates or implements learning behavior of a human to acquire new knowledge or skills, and reorganizes existing knowledge structures to continuously improve own performance. The predictive model in this implementation may be a neural network model based on machine learning,
the prediction model is a neural network model based on deep learning, which is obtained by training a large amount of sample data in advance, and comprises a plurality of layers of neural networks. In particular, the predictive model may include an input layer, an embedded layer, a sequence characterization layer, a semantic mapping layer, an expert network layer, an attention layer, a task mapping layer, and an output layer. The prediction model can also comprise a prediction network layer with a plurality of task types, so that prediction values corresponding to the plurality of task types can be calculated respectively. The predicted value may represent a predicted value of interest of the user in candidate item information.
After extracting the user behavior sequence in the user history information, the server calls a trained prediction model, inputs a plurality of candidate item information and the user behavior sequence into the prediction model, and respectively performs feature extraction on the plurality of candidate item information and the user behavior sequence through the prediction model to respectively obtain the user behavior sequence feature corresponding to the behavior type and the item feature vector of the candidate item information. The server further performs deep feature extraction and deep learning on the user behavior sequence features and the project feature vectors through the prediction model, extracts interest feature vectors and interest feature weights corresponding to the interest feature vectors, and calculates a prediction value corresponding to each candidate project information according to the target task type by using the interest feature vectors and the interest feature weights. The interest feature vector represents the interest feature vector of the user corresponding to each candidate item information.
Specifically, after the server inputs a plurality of candidate item information and a user behavior sequence into the prediction model, the data such as the candidate item information and the user behavior sequence can be preprocessed through an input layer; extracting user behavior sequence characteristics of the user behavior sequence and project characteristic vectors of candidate project information by utilizing the embedding layer; further extracting user behavior sequence features and depth features corresponding to the project feature vectors through the sequence characterization layer; the semantic mapping layer extracts semantic features of the user behavior sequence features according to the depth features corresponding to the user behavior sequence features; capturing relevance and diversity of each user behavior sequence characteristic by using an expert network layer; further improving the learning of the attention layer, extracting various combination feature information of the user behavior sequence features, and extracting an interest feature vector and an interest feature weight corresponding to the interest feature vector by combining the depth features of the project feature vector by utilizing the various combination feature information; the mapping layer further calculates the predicted value of each candidate item information according to the target task type by using the interest feature vector and the interest feature weight, and outputs the predicted value corresponding to each candidate item information through the output layer.
And step 208, screening target push information according to the predicted value of the candidate item information.
And the server calculates the predicted value of each candidate item information according to the user behavior sequence characteristic and the item characteristic vector through the prediction model, and then extracts target push information according to the predicted values corresponding to the candidate item information. Specifically, the server may screen target pushing information meeting a preset condition from the plurality of candidate item information according to the predicted value, where the server may also screen a preset number of candidate item information as target pushing information according to the preset condition. The server then generates a target push list using the extracted plurality of target push information. Wherein, the plurality refers to two or more than two, and the plurality of predicted values represents two or more than two predicted values.
Further, the server may determine target scores of the candidate item information according to the plurality of target task types by using the predicted values, so as to obtain target scores of candidate item information corresponding to the plurality of target task types, and further extract corresponding target push information according to the target scores.
Step 210, pushing the target pushing information to the user terminal corresponding to the user identifier.
After the server generates a target pushing list by utilizing the extracted multiple target pushing information, pushing the target pushing information in the target pushing list to the user terminal corresponding to the user identifier. The server can directly push the target push information to the user terminal, and can also push the target push information in a classified manner according to task types, so that the push information meeting the interests and requirements of the user can be accurately and effectively pushed to the user.
According to the information pushing method based on deep learning, after the server acquires candidate item information and the historical behavior information corresponding to the user identification, the user behavior sequences corresponding to the behavior types in the historical behavior information are extracted, so that heterogeneous behavior sequences of the user in the multiple behavior types can be effectively acquired. The candidate item information and the user behavior sequence are input into the prediction model, the item feature vector of the candidate item information and the user behavior sequence feature corresponding to the user behavior sequence are extracted, and then the prediction value of each candidate item information is determined according to the user behavior sequence feature and the item feature vector by using the prediction model based on deep learning, so that the prediction value of each candidate item information can be accurately and effectively output. Target pushing information is screened according to the predicted value of the candidate item information, the target pushing information is pushed to a user terminal corresponding to the user identifier, and the heterogeneous behavior sequence of the user and the candidate item information are learned, so that the multi-mode user interests can be accurately identified, and the accuracy of target information pushing is effectively improved.
In one embodiment, extracting a user behavior sequence corresponding to each behavior type in the historical behavior information includes: identifying a behavior type of a user behavior sequence; acquiring interest degrees corresponding to the historical behavior information; and extracting sequence features of the historical behavior information according to the interestingness to obtain a user behavior sequence corresponding to each behavior type.
Wherein the interest degree represented by each behavior type is different, for example, the interest degree of the corresponding user of the clicking behavior, the praying behavior, the comment behavior and the sharing behavior is different. The server may pre-configure a mapping table of each behavior type and interestingness. After the server inputs a plurality of user behavior sequences into the prediction model, the prediction model firstly carries out sequence feature extraction on the historical behavior information.
Specifically, the server firstly identifies the behavior types of the historical behavior information, obtains the interestingness corresponding to each behavior type according to a preset mapping relation table, and further extracts the sequence features of the historical behavior information according to the behavior types and the interestingness, so that the user behavior sequence features corresponding to a plurality of behavior types can be obtained. Sequence feature extraction is carried out on the user behavior sequence according to the behavior types and the corresponding interestingness, so that heterogeneous behavior sequences of multiple behavior types of the user can be effectively obtained.
In one embodiment, feature extraction of historical behavior information according to interest level includes: extracting a positive feedback behavior sequence and a negative feedback behavior sequence corresponding to the historical behavior information; and extracting sequence characteristics of the positive feedback behavior sequence and the negative feedback behavior sequence according to the interestingness.
The behavior types may include positive feedback behavior and negative feedback behavior, the positive feedback behavior sequence may represent a user behavior sequence of positive interest of the user, the negative feedback behavior may represent a user behavior sequence of negative interest of the user, for example, praise behavior, comment behavior, sharing behavior and the like of the user on the historical project information may be represented as positive interest of the user; the uninteresting marks of the historical project information, complaint behaviors and the like of the user can be expressed as the reverse interests of the user.
The server can extract a positive feedback behavior sequence from the historical behavior information according to the positive feedback behavior type, and extract a negative feedback behavior sequence from the user behavior sequence according to the negative feedback behavior type.
Specifically, the server identifies the positive feedback behavior type and the negative feedback behavior type of the historical behavior information according to the behavior type and the interestingness, extracts the positive feedback behavior sequence and the negative feedback behavior sequence in the historical behavior information according to the positive feedback behavior type and the negative feedback behavior type, and further extracts sequence characteristics of the positive feedback behavior sequence and the negative feedback behavior sequence according to the behavior type and the corresponding interestingness, so that user behavior sequences corresponding to all the behavior types can be generated. By identifying the positive feedback behavior type and the negative feedback behavior type of the user, the positive interest of the user can be further accurately and effectively identified, and meanwhile, the reverse interest of the user can be effectively eliminated, so that the accuracy of information extraction and pushing can be effectively improved.
In one embodiment, extracting, by the prediction model, item feature vectors corresponding to respective candidate item information includes: capturing item relevance among the candidate item information through a sequence characterization layer of the prediction model, and carrying out depth feature extraction on the candidate item information according to the item relevance to obtain item feature vectors corresponding to the candidate item information.
The sequence characterization layer of the prediction model can effectively extract each user behavior sequence and depth information contained in each candidate item information. In particular, the sequence characterization layer may be a network layer based on a transducer structure. In the process of extracting the characteristics of candidate item information through a sequence characterization layer in a prediction model, the server can capture the item relevance between each candidate item information through the sequence characterization layer based on a transducer structure, so that the depth characteristic extraction of the candidate item information can be effectively carried out according to the item relevance between each candidate item information, and further the item characteristic vector corresponding to each candidate item information is obtained.
Further, the user behavior sequence comprises behavior item information operated by the user. In the process of extracting the characteristics of the user behavior sequence by the sequence characterization layer in the server over-prediction model, the project relevance among the behavior project information can be captured by the sequence characterization layer based on the transform structure, and the deeper characteristic extraction is carried out on the user behavior sequence according to the project relevance among the behavior project information, so that the user behavior sequence characteristics corresponding to the user behavior sequences are generated. The sequence characterization layer in the prediction model is used for extracting the depth features of the user behavior sequence and the candidate item information respectively, so that the important features in the user behavior sequence and the candidate item information can be extracted accurately and effectively.
In one embodiment, as shown in fig. 3, the step of determining the predicted value of each candidate item information according to the user behavior sequence feature and the item feature vector specifically includes the following:
and 302, carrying out depth feature extraction on the user behavior sequence features to obtain a first depth feature and a second depth feature of the user behavior sequence.
And step 304, performing feature combination on the plurality of user behavior sequences according to the first depth features and the second depth features to obtain a plurality of combined feature information.
And 306, distributing corresponding combined characteristic weights to the combined characteristic information according to the target task type.
And 308, extracting interest feature vectors corresponding to the feature vectors of each item according to the combined feature information, and determining the interest feature weights of the interest feature vectors according to the combined feature weights.
Step 310, determining the predicted value of each candidate item information according to the interest feature vector and the interest feature weight.
After the server inputs the plurality of candidate item information and the user behavior sequence into the prediction model, the prediction model respectively performs feature extraction on the plurality of candidate item information and the user behavior sequence to respectively obtain item feature vectors of the user behavior sequence features and the candidate item information corresponding to the behavior types. The server further performs deep feature extraction and deep learning on the user behavior sequence features and the project feature vectors through the prediction model.
Specifically, the prediction model comprises a plurality of depth feature network layers, and the depth feature network layers are used for extracting multi-layer depth features of user behavior sequence features. For example, the prediction model includes a sequence characterization layer and a semantic mapping layer, the sequence characterization layer is used for extracting a first depth feature of the user behavior sequence feature, and the semantic mapping layer is used for further extracting a second depth feature for extracting the user behavior sequence feature according to the first depth feature. For example, the first depth feature may be a behavior-based depth feature and the second depth feature may be a semantic-based depth feature. And the prediction model further performs feature combination on the plurality of user behavior sequences according to the first depth features and the second depth features to obtain a plurality of combined feature information.
For example, the predictive model may include an expert network layer comprising a plurality of expert networks, each expert network also being comprised of a plurality of fully connected layers. And acquiring the relevance and the difference between the characteristics of each user behavior sequence according to the first depth characteristic and the second depth characteristic obtained by deep learning through an expert network layer in the prediction model, so that characteristic combination is carried out on a plurality of user behavior sequences according to the first depth characteristic and the second depth characteristic, and a plurality of combination characteristic information can be effectively obtained.
The prediction model can also comprise an attention layer, wherein the attention layer comprises a plurality of attention networks and attention weights, the attention layer can distribute corresponding combined feature weights by utilizing the combined feature information learned by the expert network layer based on different task types, and the distributed combined feature weights are taken as the attention weights of the attention layer.
After capturing a plurality of combined feature information of the user behavior sequence through an expert network layer in the prediction model, the server further extracts interest feature vectors corresponding to the feature vectors of each item through an attention layer according to the combined feature information, and determines the interest feature weights of the interest feature vectors according to the combined feature weights.
The server further determines predicted values of the candidate item information according to the interest feature vectors and the interest feature weights through the prediction model. Specifically, the server may utilize the interest feature weights to perform weighted summation on the interest feature vectors through a mapping layer in the prediction model, so as to obtain the predicted value of each candidate item information. Depth feature extraction is carried out on heterogeneous behavior sequences of users and candidate item information through a prediction model, implicit feature information of a plurality of user behavior sequences can be accurately and effectively extracted, so that predicted values of the candidate item information can be accurately calculated, and the accuracy of target information pushing can be effectively improved.
In one embodiment, feature combining the sequence of user actions according to the first depth feature and the second depth feature, obtaining a plurality of combined feature information includes: performing depth association feature extraction on the first depth feature and the second depth feature of the user behavior sequence to obtain sequence association features of the user behavior sequence; extracting sequence semantic features of the user behavior sequence according to the sequence association features; and carrying out feature combination on the plurality of user behavior sequences according to the sequence association features and the semantic features to obtain a plurality of combination feature information.
The first depth feature may be a sequence association feature of the user behavior sequence, and the second depth feature may be a sequence semantic feature of the user behavior sequence. The prediction model comprises a sequence representation layer, a semantic mapping layer and an expert network layer, wherein the sequence representation layer is used for extracting deeper features of a user behavior sequence. For example, the sequence characterization layer may be a transducer structure and adopt a residual connection mode, so that the sequence association feature of each user behavior sequence can be effectively extracted. The semantic mapping layer can be constructed in a multi-layer full-connection layer mode, and the prediction model can map the user behavior sequence features and the sequence association features into a semantic space through the semantic mapping layer so as to extract semantic features of the user behavior sequence, thereby extracting sequence semantic features of the user behavior sequence.
The prediction model further performs feature combination on the plurality of user behavior sequences according to the sequence association features and the semantic features to obtain a plurality of combination feature information, so that implicit feature information of the plurality of user behavior sequences can be accurately and effectively extracted.
In one embodiment, as shown in fig. 4, the step of combining features of multiple user behavior sequences according to the sequence relevance features and semantic features to obtain multiple combined feature information specifically includes the following steps:
step 402, performing semantic mapping on the plurality of user behavior sequence features to obtain semantic features of each user behavior sequence feature.
Step 404, mapping the plurality of user behavior sequence features to a plurality of semantic space sets according to the semantic features.
Step 406, extracting the association degree and the difference degree between the plurality of user behavior sequence features in the semantic space set according to the sequence association features, and performing feature combination on the user behavior sequence features in the semantic space set according to the association degree and the difference degree to obtain a plurality of combination feature information.
The server performs deep association feature extraction on the user behavior sequence features through the prediction model, and extracts sequence semantic features of the user behavior sequence according to the sequence association features after obtaining the sequence association features of the user behavior sequence. Specifically, the prediction model can map the user behavior sequence features and the sequence association features into a semantic space set through a semantic mapping layer so as to extract semantic features of the user behavior sequence, thereby extracting sequence semantic features of the user behavior sequence.
In order to achieve homosemantic interaction between heterogeneous behaviors of users, a semantic mapping layer can adopt a plurality of layers of full-connection layers to map sequence features and sequence association features of the behaviors of the users to a semantic space set. Specifically, the prediction model may map the user behavior sequence features into semantically identical semantic space sets according to the sequence association features. The same semantic components characterized by the heterogeneous behavior sequences can be compared in the same semantic space set.
The server extracts the association degree and the difference degree among the plurality of user behavior sequence features in the semantic space set according to the sequence association features through the prediction model, and then learns the user behavior sequence features in the semantic space set according to the association degree and the difference degree. The prediction model comprises a plurality of expert network layers, and the expert network layers are used for fully capturing the relevance and the diversity among the behavior sequence information from different angles, so that the better prediction effect can be achieved for different subsequent tasks. Each expert network is also composed of a plurality of full connection layers, different expert networks do not share network parameters of each other, and network parameters of all expert networks are shared among different tasks. Multiple expert networks are able to learn multiple combinations of heterogeneous behavior sequences of users.
And the prediction model further learns the association degree and the difference degree among the plurality of user behavior sequence features in the semantic space set through each expert network layer, and performs feature combination on the plurality of user behavior sequence features in the semantic space set according to the association degree and the difference degree, so as to obtain a plurality of combined feature information. Depth feature extraction is carried out on heterogeneous behavior sequences and candidate item information of a user through a plurality of depth feature network layers of the prediction model, so that the depth feature information of a plurality of user behavior sequences can be accurately and effectively extracted.
In one embodiment, as shown in fig. 5, there is provided an information pushing method based on deep learning, which specifically includes the following steps:
step 502, candidate item information is obtained, and historical behavior information corresponding to a user identifier is obtained.
Step 504, extracting a user behavior sequence corresponding to each behavior type in the historical behavior information.
Step 506, inputting the candidate item information and the user behavior sequence into a prediction model, and extracting the user behavior sequence characteristics corresponding to the user behavior sequence and the item characteristic vector of the candidate item information.
And step 508, extracting depth features from the user behavior sequence features to obtain a first depth feature and a second depth feature of the user behavior sequence.
And 510, performing feature combination on the plurality of user behavior sequences according to the first depth features and the second depth features to obtain a plurality of combined feature information.
And step 512, corresponding combined characteristic weights are allocated to the combined characteristic information according to the target task types.
Step 514, extracting the interest feature vectors corresponding to the feature vectors of each item according to the combined feature information, and determining the interest feature weights of the interest feature vectors according to the combined feature weights.
And step 516, inputting the interest feature vector and the interest feature weight into a target mapping layer corresponding to each target task type in the prediction model, and carrying out weighted summation on the interest feature vector by utilizing the interest feature weight according to the target task type to obtain a predicted value of the target task type corresponding to the candidate item information.
And 518, generating target scores of candidate item information according to the predicted values, and screening target pushing information meeting preset conditions from the candidate item information according to the target scores.
Step 520, pushing the target pushing information to the user terminal corresponding to the user identifier.
The prediction model comprises target mapping layers of various task types, and the target mapping layers can respectively calculate predicted values corresponding to the candidate item information and the various task types. The task types can comprise a plurality of task types such as browsing behavior prediction, praise behavior prediction, comment behavior prediction, sharing behavior prediction and the like of the candidate project information.
After extracting the interest feature weights of the interest feature vectors corresponding to the candidate item information by the user through the attention layer of the prediction model, the server inputs the interest feature vectors and the interest feature weights to the target mapping layer corresponding to each target task type in the prediction model. And each target mapping layer performs weighted summation on the interest feature vectors by utilizing the interest feature weights according to the corresponding target task types to obtain predicted values of the candidate item information corresponding to the target task types. Thus, each candidate item information can obtain one or more predicted values corresponding to the target task types.
The server may further calculate a target score for the candidate item information based on predicted values corresponding to one or more target task types of the candidate item information. The target score may be calculated according to a predicted value using a preset algorithm, for example, may be calculated by weighted summation of predicted values of a plurality of target task types using weights of the target task types, thereby obtaining a target score of each candidate item information. The server screens out target pushing information meeting preset conditions from the candidate item information according to the target scores, and then generates a target pushing list by using the extracted target pushing information, and pushes the target pushing information in the target pushing list to the user terminal corresponding to the user identifier. The predicted values of the target task types are calculated through the prediction model to push information, so that the interest degree of a user can be fully mined, and the prediction accuracy of the interest degree predicted values of the candidate item information can be effectively improved.
In one embodiment, as shown in fig. 6, before the user behavior sequence and the candidate item information are input into the prediction model, the method further comprises the step of training the prediction model, specifically comprising the following steps:
step 602, obtaining a plurality of pieces of history information, and extracting user behavior sequence data and item tags of the plurality of pieces of history information.
Step 604, generating a training set and a verification set by using the user behavior sequence data of the plurality of historical record information and the item labels.
Step 606, inputting the training set into a preset machine learning model for learning and training to obtain a training result; and iteratively updating model parameters of the machine learning model according to the training result to obtain an initial prediction model.
And 608, verifying the initial prediction model by using the verification set until the verification condition threshold is met, and obtaining the trained prediction model.
The server needs to build and train the predictive model in advance before inputting the user behavior sequence and candidate item information into the predictive model. Specifically, the server may obtain a large amount of history information of the user from the local database or the third party database in advance. For example, the server may obtain a plurality of history information from a large amount of history log information. The history record information may include labeled sample history information and unlabeled sample history information, where the labeled sample history information includes a history behavior sequence and a user interest level. The server generates a training set and a verification set using the plurality of history information. The sample history information in the training set can be marked information after manual marking, and the verification set is a plurality of unlabeled sample history information.
The server firstly carries out data cleaning and data preprocessing on the historical record information, specifically, the server carries out vectorization on the historical record information to obtain a plurality of characteristic vectors corresponding to the historical behavior sequences, and the characteristic vectors are converted into corresponding characteristic variables. The server further carries out derivative processing on the characteristic variables to obtain a plurality of processed characteristic variables. Such as filling missing values, extracting and replacing abnormal values, and the like, for the characteristic variables.
The server obtains a preset deep learning model, which may be based on a self-attention network machine learning model, for example. For example, the deep learning model includes a plurality of neural network models, and the neural network models can include a preset input layer, an embedded layer, a sequence characterization layer, a semantic mapping layer, an expert network layer, an attention network layer, a task mapping layer and an output layer. The network layer of the neural network model may include an activation function and a bias loss function. The neural network model also comprises a calculation mode for determining errors, for example, a mean square error algorithm can be adopted; the method also comprises an iterative updating mode for determining the weight parameters, for example, an Adam optimization algorithm can be adopted, and parameters in the network are updated based on training data iteration.
After the server acquires the preset deep learning model, the historical record information in the training set is input into the deep learning model for learning and training, and the training set is used for multi-target combined training of the model until the prediction effect of the model on the verification set is not improved. In the training process, there are multiple target loss functions, and the multiple target loss functions need to be fused to obtain the loss function of the whole model. For example, the loss functions of the targets may be weighted according to the directionality of the service targets, and finally the loss functions of the targets may be weighted and summed. In the training process, parameters in the network can be updated by using an Adam optimizer based on training data iteration, and a local optimal solution can be found based on an optimization problem in a reasonable time, so that a prediction model is effectively trained and continuously optimized. Whereby an initial predictive model can be trained.
After the server obtains the initial business prediction model, the historical record information in the verification set is input into the initial prediction model for further training and verification, and category probabilities corresponding to a plurality of verification data are obtained. And stopping training until the number of the condition thresholds in the verification set data reaches the verification threshold, and further obtaining a prediction model after training is completed. By training and learning a large amount of historical record information, a prediction model with high prediction accuracy can be effectively constructed and trained, so that the prediction accuracy of interest degree of candidate item information is effectively improved.
In a specific embodiment, the predictive model may include: an input layer, an embedding layer, a sequence characterization layer, a semantic mapping layer, an expert network layer, an attention layer, and task mapping and output layers of various task types. The server inputs the candidate item information and the user behavior sequence into the prediction model, and the candidate item information and the user behavior sequence are used as input data of the prediction model through an input layer, wherein the user behavior sequence comprises item information of interest of a user. The server further performs feature embedding characterization on the candidate item information and the user behavior sequence through an embedding layer in the prediction model so as to extract feature information corresponding to the candidate item information and the user behavior sequence. The embedded layer can characterize the characteristics of the candidate item information and the user behavior sequence based on the Embedding neural network, so that the characteristics of the candidate item information and the user behavior sequence can be effectively extracted. FIG. 7 is a block diagram of a predictive model, in one embodiment. Wherein, the user behavior sequence 1-the user behavior sequence N can represent a plurality of user behavior sequences of the user, and the candidate Item represents candidate Item information. As shown in fig. 7, the prediction model includes an input layer for inputting a user behavior sequence and candidate item information into a prediction model structure; the embedded layer comprises an Embedding network structure; the sequence characterization layer comprises a plurality of network layers based on a transducer structure; the semantic mapping layer comprises a plurality of layers of full-connection layers and a vector splicing layer; the expert network layer comprises a plurality of expert networks; the attention layer may include a plurality of feedforward neural networks, which may be network layers based on the Feed Forward Attention structure; the mapping layer comprises a plurality of vector splicing layers and a plurality of full-connection layers; the prediction model further includes an output layer for outputting a prediction result.
The server further performs depth feature extraction on the user behavior sequence through a sequence representation layer in the prediction model, wherein the sequence representation layer can be a structure based on a transducer, and the feature extraction is performed on the candidate item information and the user behavior sequence through the transducer structure, so that the features of the candidate item information and the user behavior sequence can be effectively extracted.
The semantic mapping layer can be a network layer connected in a full-connection layer mode, the prediction model can conduct semantic extraction and semantic interaction on the user behavior sequence features through the semantic mapping layer, and the full-connection layer is used for mapping the heterogeneous behavior sequence characterization into the same semantic space. In the same semantic space, the same semantic components of the heterogeneous behavioral sequence characterization can be compared.
The expert network layer comprises a plurality of expert networks for learning the user behavior sequence features of the same semantic space. The server can fully capture the relevance and the difference between the behavior sequence information from different angles through the expert network layer, and is beneficial to achieving better prediction effects for different subsequent tasks. Each expert network is also composed of a plurality of full-connection layers, different expert networks do not share network parameters of each other, network parameters of all the expert networks are shared among different tasks, and the expert networks can learn various combination information of heterogeneous behavior sequences of users.
The attention layer is used for calculating weights of implicit vectors output by each expert network, and carrying out weighted summation on each implicit vector by using the calculated weights to generate a characterization vector with fixed length. The equity of the combined information learned by each expert network results in information that is more relevant to the final objective not having a greater impact on the predictions, while less relevant information can produce some noise on the predictions. Thus, it is necessary to assign different importance weights to each expert network learned combination information based on different tasks. For example, the attention layer may employ a Feed Forward Attention structure to assign a respective attention weight to each expert network. The attention layer can calculate the weight of the implicit vector output by each expert network through a feedforward neural network, and uses the calculated weight to carry out weighted summation on each implicit vector and generate a characterization vector with fixed length, thereby effectively improving the prediction precision of each combined information.
The task mapping layer is used for carrying out nonlinear change on the user behavior sequence feature and the project feature vector, and different tasks do not share network parameters of the mapping layer and the output layer, so that a large amount of implicit information in the user behavior sequence feature and the project feature vector is extracted.
The output layer is used for making a final predicted value according to the extracted implicit information. The output layer in the prediction model makes final prediction according to the hidden information corresponding to the extracted target feature vector, the prediction targets of different tasks are different, and the activation functions of the final output layer are also different. For example, if the target task type is a continuous value target such as a reading time length, the activation function of the final output layer may use a ReLU function; if the target task type is a discrete value target such as clicking, the activation function of the final output layer can adopt a Sigmoid function.
In the embodiment, heterogeneous behavior sequences and candidate item information of the user are learned through the prediction model based on deep learning, so that multi-mode user interests can be accurately identified, and the accuracy of target information pushing is effectively improved.
In a specific application scenario, the information pushing method based on deep learning can fully mine long-term interests of a user by utilizing a prediction model according to historical behavior information of the user and based on heterogeneous behavior sequences in the historical behavior information of the user, and recommend most suitable target pushing information to the user. For example, the information can be pushed to the user through a social application platform, a news browsing platform, a communication platform and the like, so that personalized pushing to the user according to interest requirements of the user is realized.
Fig. 2-6 are schematic flow diagrams of a deep learning-based information pushing method in an embodiment. It should be understood that, although the steps in the flowcharts of fig. 2-6 are shown in order as indicated by the arrows, these steps are not necessarily performed in order as indicated by the arrows. The steps are not strictly limited to the order of execution unless explicitly recited herein, and the steps may be executed in other orders. Moreover, at least some of the steps in fig. 2-6 may include multiple sub-steps or stages that are not necessarily performed at the same time, but may be performed at different times, nor does the order in which the sub-steps or stages are performed necessarily occur in sequence, but may be performed alternately or alternately with at least a portion of the sub-steps or stages of other steps or other steps.
In one embodiment, as shown in fig. 8, there is provided an information pushing apparatus 800 based on deep learning, which includes an information acquisition module 802, an information prediction module 804, an information extraction module 806, and an information pushing module 808, wherein:
the information acquisition module 802 is configured to acquire candidate item information, acquire historical behavior information corresponding to a user identifier, and extract a user behavior sequence corresponding to each behavior type in the historical behavior information;
The information prediction module 804 is configured to input candidate item information and a user behavior sequence to a prediction model, extract a user behavior sequence feature corresponding to the user behavior sequence and an item feature vector of the candidate item information, and determine a predicted value of each candidate item information according to the user behavior sequence feature and the item feature vector;
the information extraction module 806 is configured to filter target push information according to the predicted value of the candidate item information;
and the information pushing module 808 is configured to push the target pushing information to the user terminal corresponding to the user identifier.
In one embodiment, the information acquisition module 802 is further configured to identify a behavior type of the user behavior sequence; acquiring the interestingness corresponding to the behavior type; and extracting features of the user behavior sequence according to the interestingness to obtain user behavior sequence features corresponding to the behavior types.
In one embodiment, the information obtaining module 802 is further configured to extract a positive feedback behavior sequence and a negative feedback behavior sequence of the user behavior sequence; and extracting features of the positive feedback behavior sequence and the negative feedback behavior sequence according to the interestingness to obtain a plurality of user behavior sequence features.
In one embodiment, the information prediction module 804 is further configured to perform depth feature extraction on the user behavior sequence feature to obtain a first depth feature and a second depth feature of the user behavior sequence; performing feature combination on the user behavior sequence according to the first depth feature and the second depth feature to obtain a plurality of combined feature information; distributing corresponding combined feature weights to the combined feature information according to the target task type; extracting interest feature vectors corresponding to the feature vectors of each item according to the combined feature information, and determining the interest feature weight of each interest feature vector according to the combined feature weight; and determining the predicted value of each candidate item information according to the interest feature vector and the interest feature weight.
In one embodiment, the information prediction module 804 is further configured to perform depth association feature extraction on the first depth feature and the second depth feature of the user behavior sequence to obtain a sequence association feature of the user behavior sequence; extracting sequence semantic features of the user behavior sequence according to the sequence association features; and carrying out feature combination on the plurality of user behavior sequences according to the sequence association features and the semantic features to obtain a plurality of combination feature information.
In one embodiment, the information prediction module 804 in the deep learning-based information push device is further configured to perform steps 402 through 406 as shown in fig. 4.
In one embodiment, the information extraction module 806 in the deep learning-based information pushing device is further configured to perform steps 506 through 516 as shown in fig. 5.
In one embodiment, as shown in fig. 9, the apparatus further includes a model training module 801, and the model training module 801 in the deep learning-based information push apparatus shown in fig. 9 is further configured to perform steps 602 to 608 shown in fig. 6.
FIG. 10 illustrates an internal block diagram of a computer device in one embodiment. The computer device may be specifically the server 104 of fig. 1. As shown in fig. 10, the computer device includes a processor, a memory, and a network interface connected by a system bus. Wherein the processor of the computer device is configured to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The database of the computer device is used for storing data such as user behavior sequences, candidate item information and the like. The network interface of the computer device is used for communicating with an external terminal through a network connection. The computer program, when executed by a processor, implements a deep learning based information push method.
It will be appreciated by those skilled in the art that the structure shown in fig. 10 is merely a block diagram of some of the structures associated with the present application and is not limiting of the computer device to which the present application may be applied, and that a particular computer device may include more or fewer components than shown, or may combine certain components, or have a different arrangement of components.
In one embodiment, the deep learning-based information pushing apparatus provided herein may be implemented in the form of a computer program that is executable on a computer device as shown in fig. 10. The memory of the computer device may store various program modules that make up the deep learning-based information pushing apparatus, such as the information acquisition module 802, the information prediction module 804, the information extraction module 806, and the information pushing module 808 shown in fig. 8. The computer program constituted by the respective program modules causes the processor to execute the steps in the deep learning-based information pushing method of the respective embodiments of the present application described in the present specification.
For example, the computer device shown in fig. 10 may perform steps 202 and 204 by the information acquisition module 802 in the deep learning-based information pushing apparatus shown in fig. 7; the computer device may perform step 206 by the information prediction module 804; the computer device may perform step 208 via the information extraction module 806; the computer device may perform step 210 via the information push module 808.
In one embodiment, a computer device is provided that includes a memory and a processor, the memory storing a computer program that, when executed by the processor, causes the processor to perform the steps of the deep learning based information push method described above. The step of the deep learning-based information pushing method herein may be a step in the deep learning-based information pushing method of the above-described respective embodiments.
In one embodiment, a computer readable storage medium is provided, storing a computer program, which when executed by a processor, causes the processor to perform the steps of the deep learning based information push method described above. The step of the deep learning-based information pushing method herein may be a step in the deep learning-based information pushing method of the above-described respective embodiments.
Those skilled in the art will appreciate that all or part of the processes in the methods of the above embodiments may be implemented by a computer program for instructing relevant hardware, where the program may be stored in a non-volatile computer readable storage medium, and where the program, when executed, may include processes in the embodiments of the methods described above. Any reference to memory, storage, database, or other medium used in the various embodiments provided herein may include non-volatile and/or volatile memory. The nonvolatile memory can include Read Only Memory (ROM), programmable ROM (PROM), electrically Programmable ROM (EPROM), electrically Erasable Programmable ROM (EEPROM), or flash memory. Volatile memory can include Random Access Memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms such as Static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double Data Rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous Link DRAM (SLDRAM), memory bus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), among others.
The technical features of the above embodiments may be arbitrarily combined, and all possible combinations of the technical features in the above embodiments are not described for brevity of description, however, as long as there is no contradiction between the combinations of the technical features, they should be considered as the scope of the description.
The above examples only represent a few embodiments of the present application, which are described in more detail and are not to be construed as limiting the scope of the present application. It should be noted that it would be apparent to those skilled in the art that various modifications and improvements could be made without departing from the spirit of the present application, which would be within the scope of the present application. Accordingly, the scope of protection of the present application is to be determined by the claims appended hereto.

Claims (14)

1. The information pushing method based on deep learning is characterized by comprising the following steps of:
candidate item information is obtained, and historical behavior information corresponding to a user identifier is obtained;
extracting a user behavior sequence corresponding to each behavior type in the historical behavior information;
inputting the candidate item information and the user behavior sequence into a prediction model, and extracting user behavior sequence characteristics corresponding to the user behavior sequence and item characteristic vectors of the candidate item information;
Depth feature extraction is carried out on the user behavior sequence features to obtain a first depth feature and a second depth feature of the user behavior sequence; performing feature combination on the user behavior sequence according to the first depth feature and the second depth feature to obtain a plurality of combined feature information; distributing corresponding combined characteristic weights to the combined characteristic information according to the target task type; extracting interest feature vectors corresponding to the feature vectors of each item according to the combined feature information, and determining the interest feature weight of each interest feature vector according to the combined feature weight; determining predicted values of candidate item information according to the interest feature vector and the interest feature weight;
screening target pushing information according to the predicted value of the candidate item information;
and pushing the target pushing information to the user terminal corresponding to the user identifier.
2. The method of claim 1, wherein the extracting the user behavior sequence corresponding to each behavior type in the historical behavior information comprises:
identifying a behavior type of the historical behavior information;
acquiring the interestingness corresponding to the behavior type;
And extracting sequence features of the historical behavior information according to the interestingness to obtain a user behavior sequence corresponding to each behavior type.
3. The method of claim 2, wherein the extracting sequence features of the historical behavior information according to the interestingness comprises:
extracting a positive feedback behavior sequence and a negative feedback behavior sequence corresponding to the historical behavior information;
and extracting sequence characteristics of the positive feedback behavior sequence and the negative feedback behavior sequence according to the interestingness.
4. The method of claim 1, wherein the feature combining the sequence of user behaviors from the first depth feature and the second depth feature to obtain a plurality of combined feature information comprises:
performing depth association feature extraction on the first depth feature and the second depth feature of the user behavior sequence to obtain a sequence association feature of the user behavior sequence;
extracting sequence semantic features of the user behavior sequence according to the sequence association features;
and carrying out feature combination on a plurality of user behavior sequences according to the sequence association features and the semantic features to obtain a plurality of combination feature information.
5. The method of claim 4, wherein the feature combining the plurality of sequences of user behavior based on the sequence-associated features and the semantic features to obtain a plurality of combined feature information comprises:
semantic mapping is carried out on the plurality of user behavior sequence features to obtain semantic features of each user behavior sequence feature;
respectively mapping the plurality of user behavior sequence features to a plurality of semantic space sets according to the semantic features;
and extracting the association degree and the difference degree among the plurality of user behavior sequence features in the semantic space set according to the sequence association features, and carrying out feature combination on the user behavior sequence features in the semantic space set according to the association degree and the difference degree to obtain a plurality of combination feature information.
6. The method according to claim 1, wherein the method further comprises:
inputting the interest feature vector and the interest feature weight into a target mapping layer corresponding to each target task type in the prediction model, and carrying out weighted summation on the interest feature vector by utilizing the interest feature weight according to the target task type to obtain a predicted value of the target task type corresponding to the candidate item information;
And generating a target score of the candidate item information by using the predicted value, and screening target pushing information meeting a preset condition from the candidate item information according to the target score.
7. An information pushing device based on deep learning, characterized in that the device comprises:
the information acquisition module is used for acquiring candidate item information, acquiring historical behavior information corresponding to a user identifier, and extracting user behavior sequences corresponding to various behavior types in the historical behavior information;
the information prediction module is used for inputting the candidate item information and the user behavior sequence into a prediction model, extracting user behavior sequence characteristics corresponding to the user behavior sequence and item characteristic vectors of the candidate item information, and carrying out depth characteristic extraction on the user behavior sequence characteristics to obtain a first depth characteristic and a second depth characteristic of the user behavior sequence; performing feature combination on the user behavior sequence according to the first depth feature and the second depth feature to obtain a plurality of combined feature information; distributing corresponding combined characteristic weights to the combined characteristic information according to the target task type; extracting interest feature vectors corresponding to the feature vectors of each item according to the combined feature information, and determining the interest feature weight of each interest feature vector according to the combined feature weight; determining predicted values of candidate item information according to the interest feature vector and the interest feature weight;
The information extraction module is used for screening target pushing information according to the predicted value of the candidate item information;
and the information pushing module is used for pushing the target pushing information to the user terminal corresponding to the user identifier.
8. The apparatus of claim 7, wherein the information acquisition module is further configured to:
identifying a behavior type of the historical behavior information;
acquiring the interestingness corresponding to the behavior type;
and extracting sequence features of the historical behavior information according to the interestingness to obtain a user behavior sequence corresponding to each behavior type.
9. The apparatus of claim 8, wherein the information acquisition module is further configured to:
extracting a positive feedback behavior sequence and a negative feedback behavior sequence corresponding to the historical behavior information;
and extracting sequence characteristics of the positive feedback behavior sequence and the negative feedback behavior sequence according to the interestingness.
10. The apparatus of claim 7, wherein the information prediction module is further configured to:
performing depth association feature extraction on the first depth feature and the second depth feature of the user behavior sequence to obtain a sequence association feature of the user behavior sequence;
Extracting sequence semantic features of the user behavior sequence according to the sequence association features;
and carrying out feature combination on a plurality of user behavior sequences according to the sequence association features and the semantic features to obtain a plurality of combination feature information.
11. The apparatus of claim 10, wherein the information prediction module is further configured to:
semantic mapping is carried out on the plurality of user behavior sequence features to obtain semantic features of each user behavior sequence feature;
respectively mapping the plurality of user behavior sequence features to a plurality of semantic space sets according to the semantic features;
and extracting the association degree and the difference degree among the plurality of user behavior sequence features in the semantic space set according to the sequence association features, and carrying out feature combination on the user behavior sequence features in the semantic space set according to the association degree and the difference degree to obtain a plurality of combination feature information.
12. The apparatus of claim 7, wherein the apparatus is further configured to:
inputting the interest feature vector and the interest feature weight into a target mapping layer corresponding to each target task type in the prediction model, and carrying out weighted summation on the interest feature vector by utilizing the interest feature weight according to the target task type to obtain a predicted value of the target task type corresponding to the candidate item information;
And generating a target score of the candidate item information by using the predicted value, and screening target pushing information meeting a preset condition from the candidate item information according to the target score.
13. A computer device comprising a memory and a processor, the memory storing a computer program, characterized in that the processor implements the steps of the method of any of claims 1 to 6 when the computer program is executed.
14. A computer readable storage medium storing a computer program, characterized in that the computer program when executed by a processor implements the steps of the method of any one of claims 1 to 6.
CN202010181488.8A 2020-03-16 2020-03-16 Information pushing method and device based on deep learning and computer equipment Active CN112905876B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202010181488.8A CN112905876B (en) 2020-03-16 2020-03-16 Information pushing method and device based on deep learning and computer equipment

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202010181488.8A CN112905876B (en) 2020-03-16 2020-03-16 Information pushing method and device based on deep learning and computer equipment

Publications (2)

Publication Number Publication Date
CN112905876A CN112905876A (en) 2021-06-04
CN112905876B true CN112905876B (en) 2024-03-15

Family

ID=76110888

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202010181488.8A Active CN112905876B (en) 2020-03-16 2020-03-16 Information pushing method and device based on deep learning and computer equipment

Country Status (1)

Country Link
CN (1) CN112905876B (en)

Families Citing this family (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113268633B (en) * 2021-06-25 2022-11-11 北京邮电大学 Short video recommendation method
CN113190599B (en) * 2021-06-30 2021-09-28 平安科技（深圳）有限公司 Processing method, device and equipment for application user behavior data and storage medium
CN113468434B (en) * 2021-09-06 2021-12-24 北京搜狐新动力信息技术有限公司 Resource recommendation method, device, readable medium and equipment
CN115774810A (en) * 2021-09-07 2023-03-10 天翼电子商务有限公司 Feature combination recommendation algorithm framework fused with sequence information
CN114780862B (en) * 2022-06-21 2022-09-23 达而观数据(成都)有限公司 User interest vector extraction method, extraction model and computer system

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108804619A (en) * 2018-05-31 2018-11-13 腾讯科技（深圳）有限公司 Interest preference prediction technique, device, computer equipment and storage medium
CN109460514A (en) * 2018-11-02 2019-03-12 北京京东尚科信息技术有限公司 Method and apparatus for pushed information
CN110413893A (en) * 2019-07-31 2019-11-05 腾讯科技（北京）有限公司 Object Push method, apparatus, computer equipment and storage medium
CN110825957A (en) * 2019-09-17 2020-02-21 中国平安人寿保险股份有限公司 Deep learning-based information recommendation method, device, equipment and storage medium

Family Cites Families (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US10757218B2 (en) * 2017-03-29 2020-08-25 Alibaba Group Holding Limited Method and apparatus for generating push notifications

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108804619A (en) * 2018-05-31 2018-11-13 腾讯科技（深圳）有限公司 Interest preference prediction technique, device, computer equipment and storage medium
CN109460514A (en) * 2018-11-02 2019-03-12 北京京东尚科信息技术有限公司 Method and apparatus for pushed information
CN110413893A (en) * 2019-07-31 2019-11-05 腾讯科技（北京）有限公司 Object Push method, apparatus, computer equipment and storage medium
CN110825957A (en) * 2019-09-17 2020-02-21 中国平安人寿保险股份有限公司 Deep learning-based information recommendation method, device, equipment and storage medium

Also Published As

Publication number Publication date
CN112905876A (en) 2021-06-04

Similar Documents

Publication Publication Date Title
CN112905876B (en) Information pushing method and device based on deep learning and computer equipment
CN110717098B (en) Meta-path-based context-aware user modeling method and sequence recommendation method
CN111966914B (en) Content recommendation method and device based on artificial intelligence and computer equipment
CN111651671B (en) User object recommendation method, device, computer equipment and storage medium
CN111859160B (en) Session sequence recommendation method and system based on graph neural network
CN109376237A (en) Prediction technique, device, computer equipment and the storage medium of client's stability
CN111291264A (en) Access object prediction method and device based on machine learning and computer equipment
CN112035611B (en) Target user recommendation method, device, computer equipment and storage medium
CN111506820A (en) Recommendation model, method, device, equipment and storage medium
CN112085541A (en) User demand analysis method and device based on browsing consumption time series data
CN111506821A (en) Recommendation model, method, device, equipment and storage medium
CN112989179B (en) Model training and multimedia content recommendation method and device
CN115063035A (en) Customer evaluation method, system, equipment and storage medium based on neural network
CN113742561A (en) Video recommendation method and device, electronic equipment and storage medium
Arora et al. Ask it right! identifying low-quality questions on community question answering services
CN116956183A (en) Multimedia resource recommendation method, model training method, device and storage medium
US11605375B2 (en) Systems and methods for dynamically updating machine learning models that provide conversational responses
US20230267277A1 (en) Systems and methods for using document activity logs to train machine-learned models for determining document relevance
KR20220152256A (en) Method and system for processing data with various temporal characteristics to generate predictions related to management agreements using random forest classifiers
US11847597B2 (en) Computer network with a performance engineering maturity model system
CN113792163B (en) Multimedia recommendation method and device, electronic equipment and storage medium
Shenouda et al. Improving Bug Assignment and Developer Allocation in Soft-ware Engineering through Interpretable Machine Learning Models
Bakanach et al. Influence of Information Technologies on Production Efficiency: Estimation on the Basis of Algorithms for Machine Learning
US20240241872A1 (en) Systems and methods for maintaining rights management while labeling data for artificial intelligence model development
US20240242115A1 (en) Systems and methods for monitoring feature engineering workflows while labeling data for artificial intelligence model development

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
REG Reference to a national code

Ref country code: HK

Ref legal event code: DE

Ref document number: 40050121

Country of ref document: HK

SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant